Fact: UNESCO guidance calls for privacy protection, age-appropriate use, human agency, teacher capacity, and institutional policy before generative AI is normalized in education.
Baseline: Schools already govern calculators, search, tutoring, collaboration, and take-home work through rules and assessment design.
Evidence conclusion: The evidence proves unmanaged AI use is risky; it does not prove all AI-supported learning is cognitive surrender.
Source: Guidance for generative AI in education and research
Fact: The U.S. Department of Education report recommends keeping humans in the loop and making AI systems aligned, inspectable, explainable, and overridable.
Baseline: That is the same control logic schools use for other educational technology: tools support learning goals instead of replacing judgment.
Evidence conclusion: The guidance supports accountable classroom design because answer substitution, privacy, and weak oversight can create real learning and governance harms.
Source: Artificial Intelligence and the Future of Teaching and Learning
Fact: Both UNESCO and the U.S. Department of Education emphasize privacy, bias, equity, teacher support, and learning design rather than treating AI use as a single yes-or-no question.
Baseline: The useful comparison is between unstructured copy-paste use and guided process evidence, revision, explanation, and defense of work.
Evidence conclusion: The conclusive point is that policy and assessment matter. If one prompt can replace the assignment, the assignment also needs evidence of thinking.
Source: UNESCO and U.S. Department of Education guidance
Fact: A 2025 integrative review synthesized 124 AI-literacy studies since 2020 and found that AI literacy spans functional, critical, and indirectly beneficial goals across technical, tool, and sociocultural perspectives.
Baseline: That is a learning-design baseline, not a ban-or-surrender baseline. Students need ways to evaluate, question, and use AI responsibly.
Evidence conclusion: The evidence supports teaching AI literacy and assessment design rather than assuming AI exposure automatically makes students stop thinking.
Source: AI Literacy in K-12 and Higher Education in the Wake of Generative AI
Fact: One preregistered trial studied nearly 1,000 high-school math students. A standard GPT-4 interface improved assisted practice by 48% but hurt later unassisted learning. A teacher-designed GPT tutor improved assisted practice by 127% and largely avoided that harm.
Baseline: The same model produced different outcomes when the interface either handed over answers or enforced tutoring guardrails.
Evidence conclusion: This directly supports the concern about unstructured answer-copying, while rejecting the claim that any AI use necessarily reduces learning.
Source: Generative AI without guardrails can harm learning
Fact: A 2025 randomized crossover trial with 194 Harvard physics students found median post-test scores of 4.5 with a research-designed AI tutor versus 3.5 after in-class active learning, from a combined pre-test median of 2.75.
Baseline: This was a purpose-built, self-paced tutor following learning-science practices, not an unrestricted chatbot writing assignments for students.
Evidence conclusion: The trial shows that designed tutoring can improve measured learning in a specific course; it does not prove every AI tutor or subject will reproduce the result.
Source: AI tutoring outperforms in-class active learning
